File size: 2,160 Bytes
1f46af1
 
acd049c
 
 
1f46af1
acd049c
1f46af1
 
 
acd049c
 
 
 
 
 
122284c
 
 
 
 
 
 
 
 
acd049c
122284c
acd049c
122284c
acd049c
122284c
 
acd049c
122284c
 
acd049c
 
122284c
 
 
 
 
acd049c
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
---
title: Eleusis
emoji: "πŸƒ"
colorFrom: green
colorTo: gray
sdk: static
app_file: index.html
pinned: false
---

# Eleusis

Play [Eleusis](https://en.wikipedia.org/wiki/Eleusis_(card_game)) β€” the card game of
scientific induction β€” against a hidden rule, and compare your game with a trained
4B model that played the **exact same deal**.

Two variants, switchable at the top of the page:

| | πŸ† Benchmark rules | πŸŽ“ Simple rules |
|---|---|---|
| Hidden rules | 26 eval rules from the [Eleusis benchmark](https://huggingface.co/datasets/nph4rd/eleusis-hf-rules) | 6 held-out simple rules Γ— 4 deals |
| Turns | 10 | 8 |
| Redraws | played card replaced from the deck | none β€” your hand shrinks |
| Model replays | [Qwen3.5-4B-eleusis-rl4-consistency-step30](https://huggingface.co/nph4rd/Qwen3.5-4B-eleusis-rl4-consistency-step30) (26Γ—8 verification run) | Qwen3.5-4B LoRA, step 90 of [this hosted RL run](https://app.primeintellect.ai/dashboard/training/r9gjcjph7qwgq0c44byyyfg9) (24Γ—8 eval) |

## How it works
1. **New game** draws a hidden rule from the selected variant's eval set.
2. Each turn you pick a card from your hand to test: the rule accepts it onto the
   mainline or rejects it.
3. At any turn you may also write a **rule guess** (a small Python predicate over
   `card` / `mainline`), or describe it in English and let ✨ Translate draft the
   Python. Guessing is free and judged by *behavioral equivalence* β€”
   the exact check used to evaluate the model.
4. **Reveal the model's game** whenever you like: real recorded games on the same
   rule and the same deal (starter card, hand, and draw order) β€” not live inference.

## Fairness notes
- You and the model face the identical deal. In the benchmark variant, if you draw
  past the model's recorded sequence, remaining draws come from a per-rule
  deterministic shuffle.
- The model submitted a rule hypothesis every turn with no tools beyond
  `play(rule, card)` β€” the same interface you have here.

Built with the [eleusis environment](https://app.primeintellect.ai/dashboard/environments/nph4rd/eleusis);
the in-Space rule engine is extracted verbatim from it.